Researchers at a multi-institutional team led by Aaqib Majeed and colleagues have developed a hybrid computational framework that combines a boundary value problem solver with an artificial neural network to analyze complex fluid dynamics. Published in Discover Nano in November 2026, the study investigates magnetohydrodynamic heat transfer in non-Newtonian Casson nanoliquid flowing across a stretching cylinder. The model incorporates the collective effects of thermal radiation, activation energy, and nth-order chemical reactions, aiming to provide a reliable tool for industrial thermal engineering applications.
The mathematical model addresses a steady two-dimensional boundary layer flow involving an incompressible nanofluid stretched at a uniform velocity. The cylinder is subjected to a radial magnetic field, and the fluid contains magnetic nanoparticles. The governing equations account for convective boundary conditions, Brownian motion, and thermophoresis to capture the transport phenomena adequately. This setup is applicable to industrial processes involving the cooling of cylindrical components and high-temperature manufacturing systems where precise control of heat and mass transfer is critical.
To solve the transformed governing differential equations, the team employed a hybrid approach using the bvp4c numerical solver and an artificial neural network. The bvp4c solver in MATLAB was used to generate a numerical dataset by solving the nonlinear ordinary differential equations derived from the similarity transformation. The solver was configured with relative tolerances of 10 to the power of negative 8 and absolute tolerances of 10 to the power of negative 10 to ensure precision in the numerical solutions.
Using the dataset generated from the bvp4c solutions, the researchers developed a Levenberg–Marquardt-based artificial neural network to improve predictive capability and computational efficiency. The dataset was subdivided into 70 percent for training, 15 percent for validation, and 15 percent for testing. The neural network architecture consisted of three input neurons, twenty hidden layer neurons, and three output neurons, utilizing tansig activation functions in the hidden layer and purelin in the output layer.
The ANN predictions exhibited excellent agreement with the numerical results, demonstrating high accuracy and strong generalization performance. The mean squared error dropped significantly to as low as 3.1901 times 10 to the power of negative 9 after 150 epochs. Regression plots for training, validation, and testing showed correlation coefficients approaching 1, indicating a near-perfect match between predicted and actual values. This validation confirms the reliability of the hybrid bvp4c–ANN framework for analyzing complex nonlinear nanofluid transport problems.
The study examined the effects of various governing parameters on velocity, temperature, and concentration distributions. The results revealed that increasing the magnetic parameter and the Casson fluid parameter suppresses the velocity field. The magnetic field creates a Lorentz force that opposes the flow of the electrically conducting fluid, while the Casson parameter increases the effective viscosity, both leading to a reduction in fluid momentum and boundary layer thickness.
Regarding thermal distribution, the findings indicated that thermal radiation and convective heating significantly enhance the temperature profile. The Eckert number, which represents the ratio of kinetic energy to enthalpy difference, was found to be in direct correlation with temperature. Higher values of the Eckert number transform more kinetic energy into internal thermal energy through frictional heating, thereby increasing the thickness of the thermal boundary layer and the overall fluid temperature.
The concentration profile was influenced by activation energy and the order of the chemical reaction. Higher activation energy intensifies the concentration profile, whereas enhancing the order of the chemical reaction reduces nanoparticle concentration. The Brownian motion parameter was observed to decrease the concentration profile as it encourages the dispersion and redistribution of nanoparticles away from the near-wall area, despite its role in enhancing the temperature distribution through improved energy exchange.
Thermophoresis, the movement of nanoparticles from hot to cold regions, was found to increase both temperature and concentration profiles. As the thermophoresis parameter increased, more nanoparticles were transported from the heat source area into the fluid, thickening the concentration layer and enhancing thermal diffusion. The curvature parameter also played a significant role, with higher curvature leading to increased velocity, temperature, and concentration due to reduced friction and enhanced uniform thermal energy distribution.
The team also quantified the variations in the wall friction coefficient, local Nusselt number, and local Sherwood number. The results showed that certain parameters had a positive influence on the skin friction coefficient, indicating greater wall shear stress, while others decreased it. Similarly, the Nusselt number increased for some parameters, signifying better heat transfer, while the Sherwood number varied based on the strength of the concentration gradient. These metrics provide a detailed understanding of the momentum, heat, and mass transfer characteristics of the flow.
The study acknowledges limitations inherent in the use of similarity transformations and numerical solutions based on the ANN-BVP4C method. These approaches may be restricted to relatively simple geometries and flows, potentially limiting their application to more complex geometries, non-similar flows, and boundary conditions. The authors suggest that future research could expand the model to include temperature- and concentration-dependent properties, particle-fluid interactions, and hybrid or ternary nanofluids in porous media.
This research contributes to the field by providing a consistent mathematical model that incorporates double stratification, velocity, and concentration slip effects alongside MHD and non-Newtonian rheology. The hybrid computational strategy offers an efficient and reliable tool for engineers and scientists working on advanced thermal engineering, energy systems, and industrial heat transfer applications. By demonstrating the coupled impact of various physical parameters, the study sheds light on the characteristics of momentum, heat, and mass transfer in Casson nanofluid systems.
The findings have practical implications for the design of heat exchangers, cooling systems, and other heat transfer equipment. The ability to tune physical parameters to control fluid velocity, heat-transfer rate, and nanoparticle concentration can lead to more efficient non-Newtonian nanofluid-based thermal management systems. The study provides a valuable theoretical basis for optimizing these systems, ensuring that the complex interactions between magnetic fields, chemical reactions, and fluid dynamics are accurately modeled and understood.
Beyond immediate industrial manufacturing applications, the research addresses longstanding thermodynamic challenges associated with conventional thermal transport media. Traditional liquids such as water, oil, and ethylene glycol exhibit relatively low thermal conductivity, which has historically constrained cooling efficiency in energy systems. The introduction of nanotechnology into fluid mechanics resolved this issue by suspending sub-100-nanometer particles composed of metals, polymers, metal oxides, or silica into the base liquid. These engineered suspensions demonstrate substantially improved thermal conductivity, viscosity, wetting behavior, and stability. In practical thermal engineering, such formulations support specialized operations ranging from electronics cooling and solar collectors to biomedical applications, including the enhancement of polymerase chain reaction efficiency using graphene-based nanofluids.
To establish the numerical integrity of their Casson fluid calculations before training predictive models, the authors conducted formal code validation against established benchmark solutions. The baseline hydrodynamic formulation describes an incompressible viscous fluid at an initial uniform ambient temperature, interacting with a cylindrical boundary extending at a linear velocity proportional to axial position. Heat exchange at the cylinder surface is governed by convective thermal conditions that balance the fluid convection coefficient with thermal conductivity. The team confirmed the reliability of their MATLAB bvp4c formulation by comparing calculated wall friction factors against prior canonical investigations published by Wang and Ishak. This comparative check verified that the reduction of the boundary layer equations via similarity transformations yielded robust solutions across limiting cases.
The computational architecture of the artificial neural network was configured specifically to act as an accelerated surrogate model for the differential equation solver. Generating the dataset required running the bvp4c solver across 500 distinct parameter combinations, partitioned into seventy percent training, fifteen percent validation, and fifteen percent testing subsets. The network utilized the Levenberg-Marquardt algorithm within the fitnet environment, accepting three primary physical inputs: the Casson parameter, thermal radiation parameter, and curvature parameter. In return, the three-neuron output layer predicted the complete dimensionless velocity, temperature, and nanoparticle concentration distributions across the boundary layer. The optimization routine operated under an error tolerance threshold of 10 to the power of negative 6 over a schedule capped at 150 training epochs.
Individual evaluations across each primary variable highlighted the numerical stability of the machine-learning surrogate. When testing network response to the thermal radiation parameter, the model converged smoothly at epoch 150 to an endorsement mean squared error of 2.4633 times 10 to the power of negative 7, with a final gradient of 1.2486 times 10 to the power of negative 7 and a damping parameter of 10 to the power of negative 8. Similarly, evaluation of the curvature parameter produced an optimal confirmation mean squared error of 7.4236 times 10 to the power of negative 7, maintaining a gradient of 2.1357 times 10 to the power of negative 6 without validation failure checks. These results verified that the trained neural network accurately reproduces the complex nonlinear interactions without incurring the repeated processing costs of full boundary value integration.
The numerical analysis also quantified secondary transport mechanisms that govern species migration and boundary layer development. Calculations conducted around baseline operating points—including a Prandtl number of 0.7, magnetic parameter of 0.1, and Brownian motion value of 0.3—demonstrated how mass diffusivity dictates concentration gradients. The Schmidt number, representing the ratio of momentum diffusivity to mass diffusivity, caused pronounced concentration drops as mass diffusivity diminished. Conversely, increasing the buoyancy ratio parameter, which weighs species concentration buoyancy against thermal buoyancy, supplied additional momentum that accelerated boundary flow. Together with surface Biot numbers defining convective-to-conductive ratios, these parametric evaluations illustrate how localized fluid resistance and thermal dispersion can be systematically adjusted across cylindrical geometries.
Subject of Research: Hybrid computational modeling of magnetohydrodynamic Casson nanofluid flow and heat transfer
Article Title: Artificial neural network and Bvp4c approach for mixed convective magnetized Casson nanoliquid flow and heat transportation over a stretchable cylinder under the presence of activation energy and nth order chemical reaction
Article References: Majeed, A., Zia, M. J., Saidani, T., Younis, J., Ali, P., Yildirim, Y., & Koam, A. N. A. (2026). Artificial neural network and Bvp4c approach for mixed convective magnetized Casson nanoliquid flow and heat transportation over a stretchable cylinder under the presence of activation energy and nth order chemical reaction. Discover Nano, 21(1), Article 591. https://doi.org/10.1186/s11671-026-05097-1
Image Credits: AI Generated
DOI: 10.1186/s11671-026-05097-1
Keywords: nanofluid, artificial neural network, magnetohydrodynamics, heat transfer, Casson fluid, bvp4c, thermal engineering, boundary layer, chemical reaction, Discover Nano, Artificial, neural
News Source: Denise Maddox. (November 11, 2026). Hybrid AI and Numerical Model Simulates Magnetic Nanofluid Flow. Scienmag.



